Large Language Model–Enabled Editing of Patient Audio Interviews From “This Is My Story” Conversations: Comparative Study
Bibliographic record
Abstract
Background: This Is My Story (TIMS) was started by Chaplain Elizabeth Tracey to promote a humanistic approach to medicine. Patients in the TIMS program are the subject of a guided conversation in which a chaplain interviews either the patient or their loved one. They are asked four questions to elicit clinically actionable information that has been shown to improve communication between patients and medical providers, strengthening medical providers' empathy. The original recorded conversation is edited into a condensed audio file approximately 1 minute and 15 seconds in length and placed in the electronic health record where it is easily accessible by all providers caring for the patient. Objective: TIMS is active at the Johns Hopkins Hospital and has shown value in assisting with provider empathy and communication. It is unique in using audio recordings to accomplish this purpose. As the program expands, there exists a barrier to adoption due to limited time and resources needed to manually edit audio conversations. To address this, we propose an automated solution using a large language model to create meaningful and concise audio summaries. Methods: We analyzed 24 TIMS audio interviews and created three edited versions of each: (1) expert-edited, (2) artificial intelligence (AI)-edited using a fully automated large language model pipeline, and (3) novice-edited by two medical students trained by the expert. A second expert, blinded to the editor, rated the audio interviews in a randomized order. This expert scored both the audio quality and content quality of each interview on 5-point Likert scales. We quantified transcript similarity to the expert-edited reference using lexical and semantic similarity metrics and identified omitted content relative to that same expert interview. Results: Audio quality (flow, pacing, clarity) and content quality (coherence, relevance, nuance) were each rated on 5-point Likert scales. Expert-edited interviews received the highest mean ratings for both audio quality (4.84) and content quality (4.83). Novice-edited scored moderately (3.84 audio, 3.63 content), while AI-edited scored slightly lower (3.49 audio, 3.20 content). Novice and AI edits were rated significantly lower than the expert edits (P<.001), but not significantly different from each other. AI and novice-edited interview transcripts had comparable overlap with the expert reference transcript, while qualitative review found frequent omissions of patient identity, actionable insights, and overall context in both the AI and novice-edited interviews. AI editing was fully automated and significantly reduced the editing time compared to both human editors. Conclusions: An AI-based editing pipeline can generate TIMS audio summaries with comparable content and audio quality to novice human editors with one hour of training. AI significantly reduces editing time and removes the need for manual training; with further validation, it could offer a solution to scale TIMS to a large range of health care settings.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.021 | 0.080 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.003 | 0.001 |
| Science and technology studies | 0.002 | 0.002 |
| Scholarly communication | 0.002 | 0.003 |
| Open science | 0.002 | 0.003 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.005 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".